PinnedMichael Lopez Chiesa·Jun 17The Anthropic Recall Was Never About Capability. It Was About a Safeguard.A jailbreak measures whether your controls hold, not what the model can do. The entire coverage cycle conflated the two.
Michael Lopez Chiesa·Jul 8Four AI Stories From This Weekand the Thing Each One Got FramesA response icon8A response icon8
Michael Lopez Chiesa·Jul 7Why You Don’t Need a Huge AI Model to do Easy WorkCompile the fuzzy behavior once with a big model, then run it locally on a 0.6B model. A new paper called Program-as-Weights makes the…
Michael Lopez Chiesa·Jul 1AI security in one week: the same force, three facesThree stories that look like they argue with each other. They don’t. They’re projections of the same object.
Michael Lopez Chiesa·Jun 30Why Verification Is Now Harder Than Generation in RL for CodeWhen your reward function becomes a fallible model instead of a ground-truth oracle, reward hacking stops being a bug and becomes the…A response icon1A response icon1
Michael Lopez Chiesa·Jun 24Three AI Stories This Week & Whether the Control Is RealAmazon on human oversight, export controls on a cyber model, and a startup claiming it beat the attention bottleneck. Same question under…A response icon3A response icon3
Michael Lopez Chiesa·Jun 23Why Agent Memory Should Store Facts, Not ConversationsA new paper from USTC argues the unit of memory is the wrong granularity in most LLM agent systems, and the fix is simpler than another…A response icon1A response icon1
Michael Lopez Chiesa·Jun 16What MiniMax Sparse Attention Actually Buys You (and What It Doesn’t)The 28.4x FLOP reduction is the boring part. The training recipe is what you came for.A response icon1A response icon1
Michael Lopez Chiesa·Jun 10Four AI Stories, One Capital CycleAnthropic filed for an IPO. SpaceX is reportedly chasing a $1.77T valuation with xAI welded onto the prospectus. OpenAI is still grinding…A response icon19A response icon19
Michael Lopez Chiesa·Jun 9What Agents’ Last Exam Actually Tells Us About AI Agent Capability in 2026The model matters three times more than the harness. The perception layer is undervalued. And the headline 26.2% pass rate means something…A response icon1A response icon1
